Transcript Xen and The Art of Virtualization

Xen and The Art of Virtualization
Paul Barham, Boris Dragovic, Keir
Fraser, Steven Hand, Tim Harris, Alex
Ho, Rolf Neugebauer, Ian Pratt &
Andrew Warfield
Jianmin Chen
10/19/2007
Motivation
 up to 100 virtual machine instances
Simultaneously on a modern server.
Concerns
 Variety: Allow for full-blown operating
systems of different types
 Commodity hardware & O/Ss, but requires
some source code modifications
 Isolation: Guest OS should not affect each
other
 Efficiency: Reduce overhead
Enhanced OS?
 It is possible to re-design O/Ss to enforce
performance isolation
 E.g. resource containers [OSDI’99]
 Difficult to account for all resource usage to a
process
 For instance, buffer cache is shared across
processes – a process may have a larger share of it
than others
 Page replacement algorithms generate process
interference – “crosstalk”
Classic VM: Full Virtualization
 Functionally identical to the underlying
machine
 Allow unmodified operating systems to be
hosted
 Difficult to handle sensitive but not
privileged instructions
 Not efficient to trap everything (Syscall …)
 Sometimes OS wants both real and virtual
resource information:
 Timer
Xen
 Design principles:
 Unmodified applications: essential
 Full-blown multi-task O/Ss: essential
 Paravirtualization: necessary for performance
and isolation
Xen
Xen VM interface: Memory
 Memory management
 Guest cannot install highest privilege level
segment descriptors; top end of linear
address space is not accessible
 Guest has direct (not trapped) read access to
hardware page tables; writes are trapped and
handled by the VMM
 Physical memory presented to guest is not
necessarily contiguous
Details: Memory 1
 TLB: challenging
 Software TLB can be virtualized without flushing TLB entries
between VM switches
 Hardware TLBs tagged with address space identifiers can also
be leveraged to avoid flushing TLB between switches
 x86 is hardware-managed and has no tags…
 Decisions:
 Guest O/Ss allocate and manage their own hardware page
tables with minimal involvement of Xen for better safety and
isolation
 Xen VMM exists in a 64MB section at the top of a VM’s address
space that is not accessible from the guest
Details: Memory 2
 Guest O/S has direct read access to
hardware page tables, but updates are
validated by the VMM
 Through “hypercalls” into Xen
 Also for segment descriptor tables
 VMM must ensure access to the Xen 64MB
section not allowed
 Guest O/S may “batch” update requests to
amortize cost of entering hypervisor
Xen VM interface: CPU
 CPU
 Guest runs at lower privilege than VMM
 Exception handlers must be registered with
VMM
 Fast system call handler can be serviced
without trapping to VMM
 Hardware interrupts replaced by lightweight
event notification system
 Timer interface: both real and virtual time
Details: CPU 1
 Protection: leverages availability of multiple “rings”
 Intermediate rings have not been used in practice since OS/2;
x86-specific
 An O/S written to only use rings 0 and 3 can be ported; needs to
modify kernel to run in ring 1
Details: CPU 2
 Frequent exceptions:
 Software interrupts for system calls
 Page faults
 Allow “guest” to register a ‘fast’ exception
handler for system calls that can be accessed
directly by CPU in ring 1, without switching to
ring-0/Xen
 Handler is validated before installing in hardware
exception table: To make sure nothing executed in
Ring 0 privilege.
 Doesn’t work for Page Fault
Xen VM interface: I/O
 I/O
 Virtual devices exposed as asynchronous I/O
rings to guests
 Event notification replaces interrupts
Details: I/O 1
 Xen does not emulate hardware devices
 Exposes device abstractions for simplicity
and performance
 I/O data transferred to/from guest via Xen
using shared-memory buffers
 Virtualized interrupts: light-weight event
delivery mechanism from Xen-guest
 Update a bitmap in shared memory
 Optional call-back handlers registered by O/S
Details: I/O 2
 I/O Descriptor Ring:
OS Porting Cost
 Number of lines of code modified or
added compared with original x86 code
base (excluding device drivers)
 Linux: 2995 (1.36%)
 Windows XP: 4620 (0.04%)
 Re-writing of privileged routines;
 Removing low-level system initialization
code
Control Transfer
 Guest synchronously call into VMM
 Explicit control transfer from guest O/S to
monitor
 “hypercalls”
 VMM delivers notifications to guest O/S
 E.g. data from an I/O device ready
 Asynchronous event mechanism; guest O/S
does not see hardware interrupts, only Xen
notifications
Event notification
 Pending events stored in per-domain
bitmask
 E.g. incoming network packet received
 Updated by Xen before invoking guest O/S
handler
 Xen-readable flag may be set by a domain
 To defer handling, based on time or number of
pending requests
 Analogous to interrupt disabling
Data Transfer: Descriptor Ring
 Descriptors are allocated by a domain
(guest) and accessible from Xen
 Descriptors do not contain I/O data;
instead, point to data buffers also
allocated by domain (guest)
 Facilitate zero-copy transfers of I/O data into
a domain
Network Virtualization
 Each domain has 1+ network interfaces
(VIFs)
 Each VIF has 2 I/O rings (send, receive)
 Each direction also has rules of the form
(<pattern>,<action>) that are inserted by
domain 0 (management)
 Xen models a virtual firewall+router (VFR)
to which all domain VIFs connect
Network Virtualization
 Packet transmission:
 Guest adds request to I/O ring
 Xen copies packet header, applies matching
filter rules
 E.g. change header IP source address for NAT
 No change to payload; pages with payload must
be pinned to physical memory until DMA to
physical NIC for transmission is complete
 Round-robin packet scheduler
Network Virtualization
 Packet reception:
 Xen applies pattern-matching rules to
determine destination VIF
 Guest O/S required to exchange unused
page frame for each packet received
 Xen exchanges packet buffer for page frame in
VIF’s receive ring
 If no receive frame is available, the packet is
dropped
 Avoids Xen-guest copies; requires pagealigned
receive buffers to be queued at VIF’s receive ring
Disk virtualization
 Domain0 has access to physical disks
 Currently: SCSI and IDE
 All other domains: virtual block device
(VBD)
 Created & configured by management
software at domain0
 Accessed via I/O ring mechanism
 Possible reordering by Xen based on
knowledge about disk layout
Disk virtualization
 Xen maintains translation tables for each
VBD
 Used to map requests for VBD (ID,offset) to
corresponding physical device and sector
address
 Zero-copy data transfers take place using
DMA between memory pages pinned by
requesting domain
 Scheduling: batches of requests in roundrobin fashion across domains
Evaluation
Microbenchmarks
 Stat, open, close, fork, exec, etc
 Xen shows overheads of up to 2x with respect
to native Linux
 (context switch across 16 processes; mmap latency)
 VMware shows up to 20x overheads
 (context switch; mmap latencies)
 UML shows up to 200x overheads
 Fork, exec, mmap; better than VMware in context
switches
Thank you